drivers/infiniband/ulp/isert/ib_isert.c

   1 /*******************************************************************************
   2  * This file contains iSCSI extentions for RDMA (iSER) Verbs
   3  *
   4  * (c) Copyright 2013 Datera, Inc.
   5  *
   6  * Nicholas A. Bellinger <nab@linux-iscsi.org>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  ****************************************************************************/
  18
  19 #include <linux/string.h>
  20 #include <linux/module.h>
  21 #include <linux/scatterlist.h>
  22 #include <linux/socket.h>
  23 #include <linux/in.h>
  24 #include <linux/in6.h>
  25 #include <rdma/ib_verbs.h>
  26 #include <rdma/rdma_cm.h>
  27 #include <target/target_core_base.h>
  28 #include <target/target_core_fabric.h>
  29 #include <target/iscsi/iscsi_transport.h>
  30 #include <linux/semaphore.h>
  31
  32 #include "ib_isert.h"
  33
  34 #define ISERT_MAX_CONN          8
  35 #define ISER_MAX_RX_CQ_LEN      (ISERT_QP_MAX_RECV_DTOS * ISERT_MAX_CONN)
  36 #define ISER_MAX_TX_CQ_LEN \
  37         ((ISERT_QP_MAX_REQ_DTOS + ISCSI_DEF_XMIT_CMDS_MAX) * ISERT_MAX_CONN)
  38 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_CQ_LEN + ISER_MAX_TX_CQ_LEN + \
  39                                  ISERT_MAX_CONN)
  40
  41 static int isert_debug_level;
  42 module_param_named(debug_level, isert_debug_level, int, 0644);
  43 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0 (default:0)");
  44
  45 static DEFINE_MUTEX(device_list_mutex);
  46 static LIST_HEAD(device_list);
  47 static struct workqueue_struct *isert_comp_wq;
  48 static struct workqueue_struct *isert_release_wq;
  49
  50 static int
  51 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd);
  52 static int
  53 isert_login_post_recv(struct isert_conn *isert_conn);
  54 static int
  55 isert_rdma_accept(struct isert_conn *isert_conn);
  56 struct rdma_cm_id *isert_setup_id(struct isert_np *isert_np);
  57
  58 static void isert_release_work(struct work_struct *work);
  59 static void isert_recv_done(struct ib_cq *cq, struct ib_wc *wc);
  60 static void isert_send_done(struct ib_cq *cq, struct ib_wc *wc);
  61 static void isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc);
  62 static void isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc);
  63
  64 static inline bool
  65 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  66 {
  67         return (conn->pi_support &&
  68                 cmd->prot_op != TARGET_PROT_NORMAL);
  69 }
  70
  71
  72 static void
  73 isert_qp_event_callback(struct ib_event *e, void *context)
  74 {
  75         struct isert_conn *isert_conn = context;
  76
  77         isert_err("%s (%d): conn %p\n",
  78                   ib_event_msg(e->event), e->event, isert_conn);
  79
  80         switch (e->event) {
  81         case IB_EVENT_COMM_EST:
  82                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  83                 break;
  84         case IB_EVENT_QP_LAST_WQE_REACHED:
  85                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  86                 break;
  87         default:
  88                 break;
  89         }
  90 }
  91
  92 static struct isert_comp *
  93 isert_comp_get(struct isert_conn *isert_conn)
  94 {
  95         struct isert_device *device = isert_conn->device;
  96         struct isert_comp *comp;
  97         int i, min = 0;
  98
  99         mutex_lock(&device_list_mutex);
 100         for (i = 0; i < device->comps_used; i++)
 101                 if (device->comps[i].active_qps <
 102                     device->comps[min].active_qps)
 103                         min = i;
 104         comp = &device->comps[min];
 105         comp->active_qps++;
 106         mutex_unlock(&device_list_mutex);
 107
 108         isert_info("conn %p, using comp %p min_index: %d\n",
 109                    isert_conn, comp, min);
 110
 111         return comp;
 112 }
 113
 114 static void
 115 isert_comp_put(struct isert_comp *comp)
 116 {
 117         mutex_lock(&device_list_mutex);
 118         comp->active_qps--;
 119         mutex_unlock(&device_list_mutex);
 120 }
 121
 122 static struct ib_qp *
 123 isert_create_qp(struct isert_conn *isert_conn,
 124                 struct isert_comp *comp,
 125                 struct rdma_cm_id *cma_id)
 126 {
 127         struct isert_device *device = isert_conn->device;
 128         struct ib_qp_init_attr attr;
 129         int ret;
 130
 131         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 132         attr.event_handler = isert_qp_event_callback;
 133         attr.qp_context = isert_conn;
 134         attr.send_cq = comp->cq;
 135         attr.recv_cq = comp->cq;
 136         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS + 1;
 137         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 138         attr.cap.max_rdma_ctxs = ISCSI_DEF_XMIT_CMDS_MAX;
 139         attr.cap.max_send_sge = device->ib_device->attrs.max_sge;
 140         attr.cap.max_recv_sge = 1;
 141         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 142         attr.qp_type = IB_QPT_RC;
 143         if (device->pi_capable)
 144                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 145
 146         ret = rdma_create_qp(cma_id, device->pd, &attr);
 147         if (ret) {
 148                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 149                 return ERR_PTR(ret);
 150         }
 151
 152         return cma_id->qp;
 153 }
 154
 155 static int
 156 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 157 {
 158         struct isert_comp *comp;
 159         int ret;
 160
 161         comp = isert_comp_get(isert_conn);
 162         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 163         if (IS_ERR(isert_conn->qp)) {
 164                 ret = PTR_ERR(isert_conn->qp);
 165                 goto err;
 166         }
 167
 168         return 0;
 169 err:
 170         isert_comp_put(comp);
 171         return ret;
 172 }
 173
 174 static int
 175 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 176 {
 177         struct isert_device *device = isert_conn->device;
 178         struct ib_device *ib_dev = device->ib_device;
 179         struct iser_rx_desc *rx_desc;
 180         struct ib_sge *rx_sg;
 181         u64 dma_addr;
 182         int i, j;
 183
 184         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 185                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 186         if (!isert_conn->rx_descs)
 187                 return -ENOMEM;
 188
 189         rx_desc = isert_conn->rx_descs;
 190
 191         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 192                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 193                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 194                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 195                         goto dma_map_fail;
 196
 197                 rx_desc->dma_addr = dma_addr;
 198
 199                 rx_sg = &rx_desc->rx_sg;
 200                 rx_sg->addr = rx_desc->dma_addr;
 201                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 202                 rx_sg->lkey = device->pd->local_dma_lkey;
 203                 rx_desc->rx_cqe.done = isert_recv_done;
 204         }
 205
 206         return 0;
 207
 208 dma_map_fail:
 209         rx_desc = isert_conn->rx_descs;
 210         for (j = 0; j < i; j++, rx_desc++) {
 211                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 212                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 213         }
 214         kfree(isert_conn->rx_descs);
 215         isert_conn->rx_descs = NULL;
 216         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 217         return -ENOMEM;
 218 }
 219
 220 static void
 221 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 222 {
 223         struct ib_device *ib_dev = isert_conn->device->ib_device;
 224         struct iser_rx_desc *rx_desc;
 225         int i;
 226
 227         if (!isert_conn->rx_descs)
 228                 return;
 229
 230         rx_desc = isert_conn->rx_descs;
 231         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 232                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 233                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 234         }
 235
 236         kfree(isert_conn->rx_descs);
 237         isert_conn->rx_descs = NULL;
 238 }
 239
 240 static void
 241 isert_free_comps(struct isert_device *device)
 242 {
 243         int i;
 244
 245         for (i = 0; i < device->comps_used; i++) {
 246                 struct isert_comp *comp = &device->comps[i];
 247
 248                 if (comp->cq)
 249                         ib_free_cq(comp->cq);
 250         }
 251         kfree(device->comps);
 252 }
 253
 254 static int
 255 isert_alloc_comps(struct isert_device *device)
 256 {
 257         int i, max_cqe, ret = 0;
 258
 259         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 260                                  device->ib_device->num_comp_vectors));
 261
 262         isert_info("Using %d CQs, %s supports %d vectors support "
 263                    "pi_capable %d\n",
 264                    device->comps_used, device->ib_device->name,
 265                    device->ib_device->num_comp_vectors,
 266                    device->pi_capable);
 267
 268         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 269                                 GFP_KERNEL);
 270         if (!device->comps)
 271                 return -ENOMEM;
 272
 273         max_cqe = min(ISER_MAX_CQ_LEN, device->ib_device->attrs.max_cqe);
 274
 275         for (i = 0; i < device->comps_used; i++) {
 276                 struct isert_comp *comp = &device->comps[i];
 277
 278                 comp->device = device;
 279                 comp->cq = ib_alloc_cq(device->ib_device, comp, max_cqe, i,
 280                                 IB_POLL_WORKQUEUE);
 281                 if (IS_ERR(comp->cq)) {
 282                         isert_err("Unable to allocate cq\n");
 283                         ret = PTR_ERR(comp->cq);
 284                         comp->cq = NULL;
 285                         goto out_cq;
 286                 }
 287         }
 288
 289         return 0;
 290 out_cq:
 291         isert_free_comps(device);
 292         return ret;
 293 }
 294
 295 static int
 296 isert_create_device_ib_res(struct isert_device *device)
 297 {
 298         struct ib_device *ib_dev = device->ib_device;
 299         int ret;
 300
 301         isert_dbg("devattr->max_sge: %d\n", ib_dev->attrs.max_sge);
 302         isert_dbg("devattr->max_sge_rd: %d\n", ib_dev->attrs.max_sge_rd);
 303
 304         ret = isert_alloc_comps(device);
 305         if (ret)
 306                 goto out;
 307
 308         device->pd = ib_alloc_pd(ib_dev, 0);
 309         if (IS_ERR(device->pd)) {
 310                 ret = PTR_ERR(device->pd);
 311                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 312                           device, ret);
 313                 goto out_cq;
 314         }
 315
 316         /* Check signature cap */
 317         device->pi_capable = ib_dev->attrs.device_cap_flags &
 318                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 319
 320         return 0;
 321
 322 out_cq:
 323         isert_free_comps(device);
 324 out:
 325         if (ret > 0)
 326                 ret = -EINVAL;
 327         return ret;
 328 }
 329
 330 static void
 331 isert_free_device_ib_res(struct isert_device *device)
 332 {
 333         isert_info("device %p\n", device);
 334
 335         ib_dealloc_pd(device->pd);
 336         isert_free_comps(device);
 337 }
 338
 339 static void
 340 isert_device_put(struct isert_device *device)
 341 {
 342         mutex_lock(&device_list_mutex);
 343         device->refcount--;
 344         isert_info("device %p refcount %d\n", device, device->refcount);
 345         if (!device->refcount) {
 346                 isert_free_device_ib_res(device);
 347                 list_del(&device->dev_node);
 348                 kfree(device);
 349         }
 350         mutex_unlock(&device_list_mutex);
 351 }
 352
 353 static struct isert_device *
 354 isert_device_get(struct rdma_cm_id *cma_id)
 355 {
 356         struct isert_device *device;
 357         int ret;
 358
 359         mutex_lock(&device_list_mutex);
 360         list_for_each_entry(device, &device_list, dev_node) {
 361                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 362                         device->refcount++;
 363                         isert_info("Found iser device %p refcount %d\n",
 364                                    device, device->refcount);
 365                         mutex_unlock(&device_list_mutex);
 366                         return device;
 367                 }
 368         }
 369
 370         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 371         if (!device) {
 372                 mutex_unlock(&device_list_mutex);
 373                 return ERR_PTR(-ENOMEM);
 374         }
 375
 376         INIT_LIST_HEAD(&device->dev_node);
 377
 378         device->ib_device = cma_id->device;
 379         ret = isert_create_device_ib_res(device);
 380         if (ret) {
 381                 kfree(device);
 382                 mutex_unlock(&device_list_mutex);
 383                 return ERR_PTR(ret);
 384         }
 385
 386         device->refcount++;
 387         list_add_tail(&device->dev_node, &device_list);
 388         isert_info("Created a new iser device %p refcount %d\n",
 389                    device, device->refcount);
 390         mutex_unlock(&device_list_mutex);
 391
 392         return device;
 393 }
 394
 395 static void
 396 isert_init_conn(struct isert_conn *isert_conn)
 397 {
 398         isert_conn->state = ISER_CONN_INIT;
 399         INIT_LIST_HEAD(&isert_conn->node);
 400         init_completion(&isert_conn->login_comp);
 401         init_completion(&isert_conn->login_req_comp);
 402         init_waitqueue_head(&isert_conn->rem_wait);
 403         kref_init(&isert_conn->kref);
 404         mutex_init(&isert_conn->mutex);
 405         INIT_WORK(&isert_conn->release_work, isert_release_work);
 406 }
 407
 408 static void
 409 isert_free_login_buf(struct isert_conn *isert_conn)
 410 {
 411         struct ib_device *ib_dev = isert_conn->device->ib_device;
 412
 413         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 414                             ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 415         kfree(isert_conn->login_rsp_buf);
 416
 417         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 418                             ISER_RX_PAYLOAD_SIZE,
 419                             DMA_FROM_DEVICE);
 420         kfree(isert_conn->login_req_buf);
 421 }
 422
 423 static int
 424 isert_alloc_login_buf(struct isert_conn *isert_conn,
 425                       struct ib_device *ib_dev)
 426 {
 427         int ret;
 428
 429         isert_conn->login_req_buf = kzalloc(sizeof(*isert_conn->login_req_buf),
 430                         GFP_KERNEL);
 431         if (!isert_conn->login_req_buf)
 432                 return -ENOMEM;
 433
 434         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 435                                 isert_conn->login_req_buf,
 436                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 437         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 438         if (ret) {
 439                 isert_err("login_req_dma mapping error: %d\n", ret);
 440                 isert_conn->login_req_dma = 0;
 441                 goto out_free_login_req_buf;
 442         }
 443
 444         isert_conn->login_rsp_buf = kzalloc(ISER_RX_PAYLOAD_SIZE, GFP_KERNEL);
 445         if (!isert_conn->login_rsp_buf) {
 446                 ret = -ENOMEM;
 447                 goto out_unmap_login_req_buf;
 448         }
 449
 450         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 451                                         isert_conn->login_rsp_buf,
 452                                         ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 453         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 454         if (ret) {
 455                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 456                 isert_conn->login_rsp_dma = 0;
 457                 goto out_free_login_rsp_buf;
 458         }
 459
 460         return 0;
 461
 462 out_free_login_rsp_buf:
 463         kfree(isert_conn->login_rsp_buf);
 464 out_unmap_login_req_buf:
 465         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 466                             ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 467 out_free_login_req_buf:
 468         kfree(isert_conn->login_req_buf);
 469         return ret;
 470 }
 471
 472 static void
 473 isert_set_nego_params(struct isert_conn *isert_conn,
 474                       struct rdma_conn_param *param)
 475 {
 476         struct ib_device_attr *attr = &isert_conn->device->ib_device->attrs;
 477
 478         /* Set max inflight RDMA READ requests */
 479         isert_conn->initiator_depth = min_t(u8, param->initiator_depth,
 480                                 attr->max_qp_init_rd_atom);
 481         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 482
 483         if (param->private_data) {
 484                 u8 flags = *(u8 *)param->private_data;
 485
 486                 /*
 487                  * use remote invalidation if the both initiator
 488                  * and the HCA support it
 489                  */
 490                 isert_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP) &&
 491                                           (attr->device_cap_flags &
 492                                            IB_DEVICE_MEM_MGT_EXTENSIONS);
 493                 if (isert_conn->snd_w_inv)
 494                         isert_info("Using remote invalidation\n");
 495         }
 496 }
 497
 498 static int
 499 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 500 {
 501         struct isert_np *isert_np = cma_id->context;
 502         struct iscsi_np *np = isert_np->np;
 503         struct isert_conn *isert_conn;
 504         struct isert_device *device;
 505         int ret = 0;
 506
 507         spin_lock_bh(&np->np_thread_lock);
 508         if (!np->enabled) {
 509                 spin_unlock_bh(&np->np_thread_lock);
 510                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 511                 return rdma_reject(cma_id, NULL, 0);
 512         }
 513         spin_unlock_bh(&np->np_thread_lock);
 514
 515         isert_dbg("cma_id: %p, portal: %p\n",
 516                  cma_id, cma_id->context);
 517
 518         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 519         if (!isert_conn)
 520                 return -ENOMEM;
 521
 522         isert_init_conn(isert_conn);
 523         isert_conn->cm_id = cma_id;
 524
 525         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 526         if (ret)
 527                 goto out;
 528
 529         device = isert_device_get(cma_id);
 530         if (IS_ERR(device)) {
 531                 ret = PTR_ERR(device);
 532                 goto out_rsp_dma_map;
 533         }
 534         isert_conn->device = device;
 535
 536         isert_set_nego_params(isert_conn, &event->param.conn);
 537
 538         ret = isert_conn_setup_qp(isert_conn, cma_id);
 539         if (ret)
 540                 goto out_conn_dev;
 541
 542         ret = isert_login_post_recv(isert_conn);
 543         if (ret)
 544                 goto out_conn_dev;
 545
 546         ret = isert_rdma_accept(isert_conn);
 547         if (ret)
 548                 goto out_conn_dev;
 549
 550         mutex_lock(&isert_np->mutex);
 551         list_add_tail(&isert_conn->node, &isert_np->accepted);
 552         mutex_unlock(&isert_np->mutex);
 553
 554         return 0;
 555
 556 out_conn_dev:
 557         isert_device_put(device);
 558 out_rsp_dma_map:
 559         isert_free_login_buf(isert_conn);
 560 out:
 561         kfree(isert_conn);
 562         rdma_reject(cma_id, NULL, 0);
 563         return ret;
 564 }
 565
 566 static void
 567 isert_connect_release(struct isert_conn *isert_conn)
 568 {
 569         struct isert_device *device = isert_conn->device;
 570
 571         isert_dbg("conn %p\n", isert_conn);
 572
 573         BUG_ON(!device);
 574
 575         isert_free_rx_descriptors(isert_conn);
 576         if (isert_conn->cm_id &&
 577             !isert_conn->dev_removed)
 578                 rdma_destroy_id(isert_conn->cm_id);
 579
 580         if (isert_conn->qp) {
 581                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 582
 583                 isert_comp_put(comp);
 584                 ib_destroy_qp(isert_conn->qp);
 585         }
 586
 587         if (isert_conn->login_req_buf)
 588                 isert_free_login_buf(isert_conn);
 589
 590         isert_device_put(device);
 591
 592         if (isert_conn->dev_removed)
 593                 wake_up_interruptible(&isert_conn->rem_wait);
 594         else
 595                 kfree(isert_conn);
 596 }
 597
 598 static void
 599 isert_connected_handler(struct rdma_cm_id *cma_id)
 600 {
 601         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 602         struct isert_np *isert_np = cma_id->context;
 603
 604         isert_info("conn %p\n", isert_conn);
 605
 606         mutex_lock(&isert_conn->mutex);
 607         isert_conn->state = ISER_CONN_UP;
 608         kref_get(&isert_conn->kref);
 609         mutex_unlock(&isert_conn->mutex);
 610
 611         mutex_lock(&isert_np->mutex);
 612         list_move_tail(&isert_conn->node, &isert_np->pending);
 613         mutex_unlock(&isert_np->mutex);
 614
 615         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 616         up(&isert_np->sem);
 617 }
 618
 619 static void
 620 isert_release_kref(struct kref *kref)
 621 {
 622         struct isert_conn *isert_conn = container_of(kref,
 623                                 struct isert_conn, kref);
 624
 625         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 626                    current->pid);
 627
 628         isert_connect_release(isert_conn);
 629 }
 630
 631 static void
 632 isert_put_conn(struct isert_conn *isert_conn)
 633 {
 634         kref_put(&isert_conn->kref, isert_release_kref);
 635 }
 636
 637 static void
 638 isert_handle_unbound_conn(struct isert_conn *isert_conn)
 639 {
 640         struct isert_np *isert_np = isert_conn->cm_id->context;
 641
 642         mutex_lock(&isert_np->mutex);
 643         if (!list_empty(&isert_conn->node)) {
 644                 /*
 645                  * This means iscsi doesn't know this connection
 646                  * so schedule a cleanup ourselves
 647                  */
 648                 list_del_init(&isert_conn->node);
 649                 isert_put_conn(isert_conn);
 650                 queue_work(isert_release_wq, &isert_conn->release_work);
 651         }
 652         mutex_unlock(&isert_np->mutex);
 653 }
 654
 655 /**
 656  * isert_conn_terminate() - Initiate connection termination
 657  * @isert_conn: isert connection struct
 658  *
 659  * Notes:
 660  * In case the connection state is BOUND, move state
 661  * to TEMINATING and start teardown sequence (rdma_disconnect).
 662  * In case the connection state is UP, complete flush as well.
 663  *
 664  * This routine must be called with mutex held. Thus it is
 665  * safe to call multiple times.
 666  */
 667 static void
 668 isert_conn_terminate(struct isert_conn *isert_conn)
 669 {
 670         int err;
 671
 672         if (isert_conn->state >= ISER_CONN_TERMINATING)
 673                 return;
 674
 675         isert_info("Terminating conn %p state %d\n",
 676                    isert_conn, isert_conn->state);
 677         isert_conn->state = ISER_CONN_TERMINATING;
 678         err = rdma_disconnect(isert_conn->cm_id);
 679         if (err)
 680                 isert_warn("Failed rdma_disconnect isert_conn %p\n",
 681                            isert_conn);
 682 }
 683
 684 static int
 685 isert_np_cma_handler(struct isert_np *isert_np,
 686                      enum rdma_cm_event_type event)
 687 {
 688         isert_dbg("%s (%d): isert np %p\n",
 689                   rdma_event_msg(event), event, isert_np);
 690
 691         switch (event) {
 692         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 693                 isert_np->cm_id = NULL;
 694                 break;
 695         case RDMA_CM_EVENT_ADDR_CHANGE:
 696                 isert_np->cm_id = isert_setup_id(isert_np);
 697                 if (IS_ERR(isert_np->cm_id)) {
 698                         isert_err("isert np %p setup id failed: %ld\n",
 699                                   isert_np, PTR_ERR(isert_np->cm_id));
 700                         isert_np->cm_id = NULL;
 701                 }
 702                 break;
 703         default:
 704                 isert_err("isert np %p Unexpected event %d\n",
 705                           isert_np, event);
 706         }
 707
 708         return -1;
 709 }
 710
 711 static int
 712 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 713                            enum rdma_cm_event_type event)
 714 {
 715         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 716
 717         mutex_lock(&isert_conn->mutex);
 718         switch (isert_conn->state) {
 719         case ISER_CONN_TERMINATING:
 720                 break;
 721         case ISER_CONN_UP:
 722                 isert_conn_terminate(isert_conn);
 723                 ib_drain_qp(isert_conn->qp);
 724                 isert_handle_unbound_conn(isert_conn);
 725                 break;
 726         case ISER_CONN_BOUND:
 727         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 728                 iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
 729                 break;
 730         default:
 731                 isert_warn("conn %p terminating in state %d\n",
 732                            isert_conn, isert_conn->state);
 733         }
 734         mutex_unlock(&isert_conn->mutex);
 735
 736         return 0;
 737 }
 738
 739 static int
 740 isert_connect_error(struct rdma_cm_id *cma_id)
 741 {
 742         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 743
 744         ib_drain_qp(isert_conn->qp);
 745         list_del_init(&isert_conn->node);
 746         isert_conn->cm_id = NULL;
 747         isert_put_conn(isert_conn);
 748
 749         return -1;
 750 }
 751
 752 static int
 753 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 754 {
 755         struct isert_np *isert_np = cma_id->context;
 756         struct isert_conn *isert_conn;
 757         int ret = 0;
 758
 759         isert_info("%s (%d): status %d id %p np %p\n",
 760                    rdma_event_msg(event->event), event->event,
 761                    event->status, cma_id, cma_id->context);
 762
 763         if (isert_np->cm_id == cma_id)
 764                 return isert_np_cma_handler(cma_id->context, event->event);
 765
 766         switch (event->event) {
 767         case RDMA_CM_EVENT_CONNECT_REQUEST:
 768                 ret = isert_connect_request(cma_id, event);
 769                 if (ret)
 770                         isert_err("failed handle connect request %d\n", ret);
 771                 break;
 772         case RDMA_CM_EVENT_ESTABLISHED:
 773                 isert_connected_handler(cma_id);
 774                 break;
 775         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 776         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 777         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 778                 ret = isert_disconnected_handler(cma_id, event->event);
 779                 break;
 780         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 781                 isert_conn = cma_id->qp->qp_context;
 782                 isert_conn->dev_removed = true;
 783                 isert_disconnected_handler(cma_id, event->event);
 784                 wait_event_interruptible(isert_conn->rem_wait,
 785                                          isert_conn->state == ISER_CONN_DOWN);
 786                 kfree(isert_conn);
 787                 /*
 788                  * return non-zero from the callback to destroy
 789                  * the rdma cm id
 790                  */
 791                 return 1;
 792         case RDMA_CM_EVENT_REJECTED:
 793                 isert_info("Connection rejected: %s\n",
 794                            rdma_reject_msg(cma_id, event->status));
 795                 /* fall through */
 796         case RDMA_CM_EVENT_UNREACHABLE:
 797         case RDMA_CM_EVENT_CONNECT_ERROR:
 798                 ret = isert_connect_error(cma_id);
 799                 break;
 800         default:
 801                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 802                 break;
 803         }
 804
 805         return ret;
 806 }
 807
 808 static int
 809 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 810 {
 811         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 812         int i, ret;
 813         struct iser_rx_desc *rx_desc;
 814
 815         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 816                 rx_desc = &isert_conn->rx_descs[i];
 817
 818                 rx_wr->wr_cqe = &rx_desc->rx_cqe;
 819                 rx_wr->sg_list = &rx_desc->rx_sg;
 820                 rx_wr->num_sge = 1;
 821                 rx_wr->next = rx_wr + 1;
 822                 rx_desc->in_use = false;
 823         }
 824         rx_wr--;
 825         rx_wr->next = NULL; /* mark end of work requests list */
 826
 827         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 828                            &rx_wr_failed);
 829         if (ret)
 830                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 831
 832         return ret;
 833 }
 834
 835 static int
 836 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 837 {
 838         struct ib_recv_wr *rx_wr_failed, rx_wr;
 839         int ret;
 840
 841         if (!rx_desc->in_use) {
 842                 /*
 843                  * if the descriptor is not in-use we already reposted it
 844                  * for recv, so just silently return
 845                  */
 846                 return 0;
 847         }
 848
 849         rx_desc->in_use = false;
 850         rx_wr.wr_cqe = &rx_desc->rx_cqe;
 851         rx_wr.sg_list = &rx_desc->rx_sg;
 852         rx_wr.num_sge = 1;
 853         rx_wr.next = NULL;
 854
 855         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
 856         if (ret)
 857                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 858
 859         return ret;
 860 }
 861
 862 static int
 863 isert_login_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
 864 {
 865         struct ib_device *ib_dev = isert_conn->cm_id->device;
 866         struct ib_send_wr send_wr, *send_wr_failed;
 867         int ret;
 868
 869         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
 870                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
 871
 872         tx_desc->tx_cqe.done = isert_login_send_done;
 873
 874         send_wr.next    = NULL;
 875         send_wr.wr_cqe  = &tx_desc->tx_cqe;
 876         send_wr.sg_list = tx_desc->tx_sg;
 877         send_wr.num_sge = tx_desc->num_sge;
 878         send_wr.opcode  = IB_WR_SEND;
 879         send_wr.send_flags = IB_SEND_SIGNALED;
 880
 881         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
 882         if (ret)
 883                 isert_err("ib_post_send() failed, ret: %d\n", ret);
 884
 885         return ret;
 886 }
 887
 888 static void
 889 __isert_create_send_desc(struct isert_device *device,
 890                          struct iser_tx_desc *tx_desc)
 891 {
 892
 893         memset(&tx_desc->iser_header, 0, sizeof(struct iser_ctrl));
 894         tx_desc->iser_header.flags = ISCSI_CTRL;
 895
 896         tx_desc->num_sge = 1;
 897
 898         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
 899                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 900                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
 901         }
 902 }
 903
 904 static void
 905 isert_create_send_desc(struct isert_conn *isert_conn,
 906                        struct isert_cmd *isert_cmd,
 907                        struct iser_tx_desc *tx_desc)
 908 {
 909         struct isert_device *device = isert_conn->device;
 910         struct ib_device *ib_dev = device->ib_device;
 911
 912         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
 913                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
 914
 915         __isert_create_send_desc(device, tx_desc);
 916 }
 917
 918 static int
 919 isert_init_tx_hdrs(struct isert_conn *isert_conn,
 920                    struct iser_tx_desc *tx_desc)
 921 {
 922         struct isert_device *device = isert_conn->device;
 923         struct ib_device *ib_dev = device->ib_device;
 924         u64 dma_addr;
 925
 926         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
 927                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
 928         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
 929                 isert_err("ib_dma_mapping_error() failed\n");
 930                 return -ENOMEM;
 931         }
 932
 933         tx_desc->dma_addr = dma_addr;
 934         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
 935         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
 936         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 937
 938         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
 939                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
 940                   tx_desc->tx_sg[0].lkey);
 941
 942         return 0;
 943 }
 944
 945 static void
 946 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
 947                    struct ib_send_wr *send_wr)
 948 {
 949         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
 950
 951         tx_desc->tx_cqe.done = isert_send_done;
 952         send_wr->wr_cqe = &tx_desc->tx_cqe;
 953
 954         if (isert_conn->snd_w_inv && isert_cmd->inv_rkey) {
 955                 send_wr->opcode  = IB_WR_SEND_WITH_INV;
 956                 send_wr->ex.invalidate_rkey = isert_cmd->inv_rkey;
 957         } else {
 958                 send_wr->opcode = IB_WR_SEND;
 959         }
 960
 961         send_wr->sg_list = &tx_desc->tx_sg[0];
 962         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
 963         send_wr->send_flags = IB_SEND_SIGNALED;
 964 }
 965
 966 static int
 967 isert_login_post_recv(struct isert_conn *isert_conn)
 968 {
 969         struct ib_recv_wr rx_wr, *rx_wr_fail;
 970         struct ib_sge sge;
 971         int ret;
 972
 973         memset(&sge, 0, sizeof(struct ib_sge));
 974         sge.addr = isert_conn->login_req_dma;
 975         sge.length = ISER_RX_PAYLOAD_SIZE;
 976         sge.lkey = isert_conn->device->pd->local_dma_lkey;
 977
 978         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
 979                 sge.addr, sge.length, sge.lkey);
 980
 981         isert_conn->login_req_buf->rx_cqe.done = isert_login_recv_done;
 982
 983         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
 984         rx_wr.wr_cqe = &isert_conn->login_req_buf->rx_cqe;
 985         rx_wr.sg_list = &sge;
 986         rx_wr.num_sge = 1;
 987
 988         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
 989         if (ret)
 990                 isert_err("ib_post_recv() failed: %d\n", ret);
 991
 992         return ret;
 993 }
 994
 995 static int
 996 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
 997                    u32 length)
 998 {
 999         struct isert_conn *isert_conn = conn->context;
1000         struct isert_device *device = isert_conn->device;
1001         struct ib_device *ib_dev = device->ib_device;
1002         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1003         int ret;
1004
1005         __isert_create_send_desc(device, tx_desc);
1006
1007         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1008                sizeof(struct iscsi_hdr));
1009
1010         isert_init_tx_hdrs(isert_conn, tx_desc);
1011
1012         if (length > 0) {
1013                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1014
1015                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1016                                            length, DMA_TO_DEVICE);
1017
1018                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1019
1020                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1021                                               length, DMA_TO_DEVICE);
1022
1023                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1024                 tx_dsg->length  = length;
1025                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1026                 tx_desc->num_sge = 2;
1027         }
1028         if (!login->login_failed) {
1029                 if (login->login_complete) {
1030                         ret = isert_alloc_rx_descriptors(isert_conn);
1031                         if (ret)
1032                                 return ret;
1033
1034                         ret = isert_post_recvm(isert_conn,
1035                                                ISERT_QP_MAX_RECV_DTOS);
1036                         if (ret)
1037                                 return ret;
1038
1039                         /* Now we are in FULL_FEATURE phase */
1040                         mutex_lock(&isert_conn->mutex);
1041                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1042                         mutex_unlock(&isert_conn->mutex);
1043                         goto post_send;
1044                 }
1045
1046                 ret = isert_login_post_recv(isert_conn);
1047                 if (ret)
1048                         return ret;
1049         }
1050 post_send:
1051         ret = isert_login_post_send(isert_conn, tx_desc);
1052         if (ret)
1053                 return ret;
1054
1055         return 0;
1056 }
1057
1058 static void
1059 isert_rx_login_req(struct isert_conn *isert_conn)
1060 {
1061         struct iser_rx_desc *rx_desc = isert_conn->login_req_buf;
1062         int rx_buflen = isert_conn->login_req_len;
1063         struct iscsi_conn *conn = isert_conn->conn;
1064         struct iscsi_login *login = conn->conn_login;
1065         int size;
1066
1067         isert_info("conn %p\n", isert_conn);
1068
1069         WARN_ON_ONCE(!login);
1070
1071         if (login->first_request) {
1072                 struct iscsi_login_req *login_req =
1073                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1074                 /*
1075                  * Setup the initial iscsi_login values from the leading
1076                  * login request PDU.
1077                  */
1078                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1079                 login->current_stage =
1080                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1081                          >> 2;
1082                 login->version_min      = login_req->min_version;
1083                 login->version_max      = login_req->max_version;
1084                 memcpy(login->isid, login_req->isid, 6);
1085                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1086                 login->init_task_tag    = login_req->itt;
1087                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1088                 login->cid              = be16_to_cpu(login_req->cid);
1089                 login->tsih             = be16_to_cpu(login_req->tsih);
1090         }
1091
1092         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1093
1094         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1095         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1096                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1097                   MAX_KEY_VALUE_PAIRS);
1098         memcpy(login->req_buf, &rx_desc->data[0], size);
1099
1100         if (login->first_request) {
1101                 complete(&isert_conn->login_comp);
1102                 return;
1103         }
1104         schedule_delayed_work(&conn->login_work, 0);
1105 }
1106
1107 static struct iscsi_cmd
1108 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1109 {
1110         struct isert_conn *isert_conn = conn->context;
1111         struct isert_cmd *isert_cmd;
1112         struct iscsi_cmd *cmd;
1113
1114         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1115         if (!cmd) {
1116                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1117                 return NULL;
1118         }
1119         isert_cmd = iscsit_priv_cmd(cmd);
1120         isert_cmd->conn = isert_conn;
1121         isert_cmd->iscsi_cmd = cmd;
1122         isert_cmd->rx_desc = rx_desc;
1123
1124         return cmd;
1125 }
1126
1127 static int
1128 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1129                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1130                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1131 {
1132         struct iscsi_conn *conn = isert_conn->conn;
1133         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1134         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1135         bool dump_payload = false;
1136         unsigned int data_len;
1137
1138         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1139         if (rc < 0)
1140                 return rc;
1141
1142         imm_data = cmd->immediate_data;
1143         imm_data_len = cmd->first_burst_len;
1144         unsol_data = cmd->unsolicited_data;
1145         data_len = cmd->se_cmd.data_length;
1146
1147         if (imm_data && imm_data_len == data_len)
1148                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1149         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1150         if (rc < 0) {
1151                 return 0;
1152         } else if (rc > 0) {
1153                 dump_payload = true;
1154                 goto sequence_cmd;
1155         }
1156
1157         if (!imm_data)
1158                 return 0;
1159
1160         if (imm_data_len != data_len) {
1161                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1162                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1163                                     &rx_desc->data[0], imm_data_len);
1164                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1165                           sg_nents, imm_data_len);
1166         } else {
1167                 sg_init_table(&isert_cmd->sg, 1);
1168                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1169                 cmd->se_cmd.t_data_nents = 1;
1170                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1171                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1172                           imm_data_len);
1173         }
1174
1175         cmd->write_data_done += imm_data_len;
1176
1177         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1178                 spin_lock_bh(&cmd->istate_lock);
1179                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1180                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1181                 spin_unlock_bh(&cmd->istate_lock);
1182         }
1183
1184 sequence_cmd:
1185         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1186
1187         if (!rc && dump_payload == false && unsol_data)
1188                 iscsit_set_unsoliticed_dataout(cmd);
1189         else if (dump_payload && imm_data)
1190                 target_put_sess_cmd(&cmd->se_cmd);
1191
1192         return 0;
1193 }
1194
1195 static int
1196 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1197                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1198 {
1199         struct scatterlist *sg_start;
1200         struct iscsi_conn *conn = isert_conn->conn;
1201         struct iscsi_cmd *cmd = NULL;
1202         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1203         u32 unsol_data_len = ntoh24(hdr->dlength);
1204         int rc, sg_nents, sg_off, page_off;
1205
1206         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1207         if (rc < 0)
1208                 return rc;
1209         else if (!cmd)
1210                 return 0;
1211         /*
1212          * FIXME: Unexpected unsolicited_data out
1213          */
1214         if (!cmd->unsolicited_data) {
1215                 isert_err("Received unexpected solicited data payload\n");
1216                 dump_stack();
1217                 return -1;
1218         }
1219
1220         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1221                   "write_data_done: %u, data_length: %u\n",
1222                   unsol_data_len,  cmd->write_data_done,
1223                   cmd->se_cmd.data_length);
1224
1225         sg_off = cmd->write_data_done / PAGE_SIZE;
1226         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1227         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1228         page_off = cmd->write_data_done % PAGE_SIZE;
1229         /*
1230          * FIXME: Non page-aligned unsolicited_data out
1231          */
1232         if (page_off) {
1233                 isert_err("unexpected non-page aligned data payload\n");
1234                 dump_stack();
1235                 return -1;
1236         }
1237         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1238                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1239                   sg_nents, &rx_desc->data[0], unsol_data_len);
1240
1241         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1242                             unsol_data_len);
1243
1244         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1245         if (rc < 0)
1246                 return rc;
1247
1248         /*
1249          * multiple data-outs on the same command can arrive -
1250          * so post the buffer before hand
1251          */
1252         rc = isert_post_recv(isert_conn, rx_desc);
1253         if (rc) {
1254                 isert_err("ib_post_recv failed with %d\n", rc);
1255                 return rc;
1256         }
1257         return 0;
1258 }
1259
1260 static int
1261 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1262                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1263                      unsigned char *buf)
1264 {
1265         struct iscsi_conn *conn = isert_conn->conn;
1266         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1267         int rc;
1268
1269         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1270         if (rc < 0)
1271                 return rc;
1272         /*
1273          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1274          */
1275
1276         return iscsit_process_nop_out(conn, cmd, hdr);
1277 }
1278
1279 static int
1280 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1281                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1282                       struct iscsi_text *hdr)
1283 {
1284         struct iscsi_conn *conn = isert_conn->conn;
1285         u32 payload_length = ntoh24(hdr->dlength);
1286         int rc;
1287         unsigned char *text_in = NULL;
1288
1289         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1290         if (rc < 0)
1291                 return rc;
1292
1293         if (payload_length) {
1294                 text_in = kzalloc(payload_length, GFP_KERNEL);
1295                 if (!text_in)
1296                         return -ENOMEM;
1297         }
1298         cmd->text_in_ptr = text_in;
1299
1300         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1301
1302         return iscsit_process_text_cmd(conn, cmd, hdr);
1303 }
1304
1305 static int
1306 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1307                 uint32_t read_stag, uint64_t read_va,
1308                 uint32_t write_stag, uint64_t write_va)
1309 {
1310         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1311         struct iscsi_conn *conn = isert_conn->conn;
1312         struct iscsi_cmd *cmd;
1313         struct isert_cmd *isert_cmd;
1314         int ret = -EINVAL;
1315         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1316
1317         if (conn->sess->sess_ops->SessionType &&
1318            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1319                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1320                           " ignoring\n", opcode);
1321                 return 0;
1322         }
1323
1324         switch (opcode) {
1325         case ISCSI_OP_SCSI_CMD:
1326                 cmd = isert_allocate_cmd(conn, rx_desc);
1327                 if (!cmd)
1328                         break;
1329
1330                 isert_cmd = iscsit_priv_cmd(cmd);
1331                 isert_cmd->read_stag = read_stag;
1332                 isert_cmd->read_va = read_va;
1333                 isert_cmd->write_stag = write_stag;
1334                 isert_cmd->write_va = write_va;
1335                 isert_cmd->inv_rkey = read_stag ? read_stag : write_stag;
1336
1337                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1338                                         rx_desc, (unsigned char *)hdr);
1339                 break;
1340         case ISCSI_OP_NOOP_OUT:
1341                 cmd = isert_allocate_cmd(conn, rx_desc);
1342                 if (!cmd)
1343                         break;
1344
1345                 isert_cmd = iscsit_priv_cmd(cmd);
1346                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1347                                            rx_desc, (unsigned char *)hdr);
1348                 break;
1349         case ISCSI_OP_SCSI_DATA_OUT:
1350                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1351                                                 (unsigned char *)hdr);
1352                 break;
1353         case ISCSI_OP_SCSI_TMFUNC:
1354                 cmd = isert_allocate_cmd(conn, rx_desc);
1355                 if (!cmd)
1356                         break;
1357
1358                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1359                                                 (unsigned char *)hdr);
1360                 break;
1361         case ISCSI_OP_LOGOUT:
1362                 cmd = isert_allocate_cmd(conn, rx_desc);
1363                 if (!cmd)
1364                         break;
1365
1366                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1367                 break;
1368         case ISCSI_OP_TEXT:
1369                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1370                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1371                 else
1372                         cmd = isert_allocate_cmd(conn, rx_desc);
1373
1374                 if (!cmd)
1375                         break;
1376
1377                 isert_cmd = iscsit_priv_cmd(cmd);
1378                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1379                                             rx_desc, (struct iscsi_text *)hdr);
1380                 break;
1381         default:
1382                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1383                 dump_stack();
1384                 break;
1385         }
1386
1387         return ret;
1388 }
1389
1390 static void
1391 isert_print_wc(struct ib_wc *wc, const char *type)
1392 {
1393         if (wc->status != IB_WC_WR_FLUSH_ERR)
1394                 isert_err("%s failure: %s (%d) vend_err %x\n", type,
1395                           ib_wc_status_msg(wc->status), wc->status,
1396                           wc->vendor_err);
1397         else
1398                 isert_dbg("%s failure: %s (%d)\n", type,
1399                           ib_wc_status_msg(wc->status), wc->status);
1400 }
1401
1402 static void
1403 isert_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1404 {
1405         struct isert_conn *isert_conn = wc->qp->qp_context;
1406         struct ib_device *ib_dev = isert_conn->cm_id->device;
1407         struct iser_rx_desc *rx_desc = cqe_to_rx_desc(wc->wr_cqe);
1408         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1409         struct iser_ctrl *iser_ctrl = &rx_desc->iser_header;
1410         uint64_t read_va = 0, write_va = 0;
1411         uint32_t read_stag = 0, write_stag = 0;
1412
1413         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1414                 isert_print_wc(wc, "recv");
1415                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1416                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1417                 return;
1418         }
1419
1420         rx_desc->in_use = true;
1421
1422         ib_dma_sync_single_for_cpu(ib_dev, rx_desc->dma_addr,
1423                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1424
1425         isert_dbg("DMA: 0x%llx, iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1426                  rx_desc->dma_addr, hdr->opcode, hdr->itt, hdr->flags,
1427                  (int)(wc->byte_len - ISER_HEADERS_LEN));
1428
1429         switch (iser_ctrl->flags & 0xF0) {
1430         case ISCSI_CTRL:
1431                 if (iser_ctrl->flags & ISER_RSV) {
1432                         read_stag = be32_to_cpu(iser_ctrl->read_stag);
1433                         read_va = be64_to_cpu(iser_ctrl->read_va);
1434                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1435                                   read_stag, (unsigned long long)read_va);
1436                 }
1437                 if (iser_ctrl->flags & ISER_WSV) {
1438                         write_stag = be32_to_cpu(iser_ctrl->write_stag);
1439                         write_va = be64_to_cpu(iser_ctrl->write_va);
1440                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1441                                   write_stag, (unsigned long long)write_va);
1442                 }
1443
1444                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1445                 break;
1446         case ISER_HELLO:
1447                 isert_err("iSER Hello message\n");
1448                 break;
1449         default:
1450                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_ctrl->flags);
1451                 break;
1452         }
1453
1454         isert_rx_opcode(isert_conn, rx_desc,
1455                         read_stag, read_va, write_stag, write_va);
1456
1457         ib_dma_sync_single_for_device(ib_dev, rx_desc->dma_addr,
1458                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1459 }
1460
1461 static void
1462 isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1463 {
1464         struct isert_conn *isert_conn = wc->qp->qp_context;
1465         struct ib_device *ib_dev = isert_conn->device->ib_device;
1466
1467         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1468                 isert_print_wc(wc, "login recv");
1469                 return;
1470         }
1471
1472         ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_req_dma,
1473                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1474
1475         isert_conn->login_req_len = wc->byte_len - ISER_HEADERS_LEN;
1476
1477         if (isert_conn->conn) {
1478                 struct iscsi_login *login = isert_conn->conn->conn_login;
1479
1480                 if (login && !login->first_request)
1481                         isert_rx_login_req(isert_conn);
1482         }
1483
1484         mutex_lock(&isert_conn->mutex);
1485         complete(&isert_conn->login_req_comp);
1486         mutex_unlock(&isert_conn->mutex);
1487
1488         ib_dma_sync_single_for_device(ib_dev, isert_conn->login_req_dma,
1489                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1490 }
1491
1492 static void
1493 isert_rdma_rw_ctx_destroy(struct isert_cmd *cmd, struct isert_conn *conn)
1494 {
1495         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
1496         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
1497
1498         if (!cmd->rw.nr_ops)
1499                 return;
1500
1501         if (isert_prot_cmd(conn, se_cmd)) {
1502                 rdma_rw_ctx_destroy_signature(&cmd->rw, conn->qp,
1503                                 conn->cm_id->port_num, se_cmd->t_data_sg,
1504                                 se_cmd->t_data_nents, se_cmd->t_prot_sg,
1505                                 se_cmd->t_prot_nents, dir);
1506         } else {
1507                 rdma_rw_ctx_destroy(&cmd->rw, conn->qp, conn->cm_id->port_num,
1508                                 se_cmd->t_data_sg, se_cmd->t_data_nents, dir);
1509         }
1510
1511         cmd->rw.nr_ops = 0;
1512 }
1513
1514 static void
1515 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1516 {
1517         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1518         struct isert_conn *isert_conn = isert_cmd->conn;
1519         struct iscsi_conn *conn = isert_conn->conn;
1520         struct iscsi_text_rsp *hdr;
1521
1522         isert_dbg("Cmd %p\n", isert_cmd);
1523
1524         switch (cmd->iscsi_opcode) {
1525         case ISCSI_OP_SCSI_CMD:
1526                 spin_lock_bh(&conn->cmd_lock);
1527                 if (!list_empty(&cmd->i_conn_node))
1528                         list_del_init(&cmd->i_conn_node);
1529                 spin_unlock_bh(&conn->cmd_lock);
1530
1531                 if (cmd->data_direction == DMA_TO_DEVICE) {
1532                         iscsit_stop_dataout_timer(cmd);
1533                         /*
1534                          * Check for special case during comp_err where
1535                          * WRITE_PENDING has been handed off from core,
1536                          * but requires an extra target_put_sess_cmd()
1537                          * before transport_generic_free_cmd() below.
1538                          */
1539                         if (comp_err &&
1540                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1541                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1542
1543                                 target_put_sess_cmd(se_cmd);
1544                         }
1545                 }
1546
1547                 isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1548                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1549                 break;
1550         case ISCSI_OP_SCSI_TMFUNC:
1551                 spin_lock_bh(&conn->cmd_lock);
1552                 if (!list_empty(&cmd->i_conn_node))
1553                         list_del_init(&cmd->i_conn_node);
1554                 spin_unlock_bh(&conn->cmd_lock);
1555
1556                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1557                 break;
1558         case ISCSI_OP_REJECT:
1559         case ISCSI_OP_NOOP_OUT:
1560         case ISCSI_OP_TEXT:
1561                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1562                 /* If the continue bit is on, keep the command alive */
1563                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1564                         break;
1565
1566                 spin_lock_bh(&conn->cmd_lock);
1567                 if (!list_empty(&cmd->i_conn_node))
1568                         list_del_init(&cmd->i_conn_node);
1569                 spin_unlock_bh(&conn->cmd_lock);
1570
1571                 /*
1572                  * Handle special case for REJECT when iscsi_add_reject*() has
1573                  * overwritten the original iscsi_opcode assignment, and the
1574                  * associated cmd->se_cmd needs to be released.
1575                  */
1576                 if (cmd->se_cmd.se_tfo != NULL) {
1577                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1578                                  cmd->iscsi_opcode);
1579                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1580                         break;
1581                 }
1582                 /* fall through */
1583         default:
1584                 iscsit_release_cmd(cmd);
1585                 break;
1586         }
1587 }
1588
1589 static void
1590 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1591 {
1592         if (tx_desc->dma_addr != 0) {
1593                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1594                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1595                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1596                 tx_desc->dma_addr = 0;
1597         }
1598 }
1599
1600 static void
1601 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1602                      struct ib_device *ib_dev, bool comp_err)
1603 {
1604         if (isert_cmd->pdu_buf_dma != 0) {
1605                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1606                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1607                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1608                 isert_cmd->pdu_buf_dma = 0;
1609         }
1610
1611         isert_unmap_tx_desc(tx_desc, ib_dev);
1612         isert_put_cmd(isert_cmd, comp_err);
1613 }
1614
1615 static int
1616 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1617 {
1618         struct ib_mr_status mr_status;
1619         int ret;
1620
1621         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1622         if (ret) {
1623                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1624                 goto fail_mr_status;
1625         }
1626
1627         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1628                 u64 sec_offset_err;
1629                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1630
1631                 switch (mr_status.sig_err.err_type) {
1632                 case IB_SIG_BAD_GUARD:
1633                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1634                         break;
1635                 case IB_SIG_BAD_REFTAG:
1636                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1637                         break;
1638                 case IB_SIG_BAD_APPTAG:
1639                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1640                         break;
1641                 }
1642                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1643                 do_div(sec_offset_err, block_size);
1644                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1645
1646                 isert_err("PI error found type %d at sector 0x%llx "
1647                           "expected 0x%x vs actual 0x%x\n",
1648                           mr_status.sig_err.err_type,
1649                           (unsigned long long)se_cmd->bad_sector,
1650                           mr_status.sig_err.expected,
1651                           mr_status.sig_err.actual);
1652                 ret = 1;
1653         }
1654
1655 fail_mr_status:
1656         return ret;
1657 }
1658
1659 static void
1660 isert_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
1661 {
1662         struct isert_conn *isert_conn = wc->qp->qp_context;
1663         struct isert_device *device = isert_conn->device;
1664         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1665         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1666         struct se_cmd *cmd = &isert_cmd->iscsi_cmd->se_cmd;
1667         int ret = 0;
1668
1669         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1670                 isert_print_wc(wc, "rdma write");
1671                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1672                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1673                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1674                 return;
1675         }
1676
1677         isert_dbg("Cmd %p\n", isert_cmd);
1678
1679         ret = isert_check_pi_status(cmd, isert_cmd->rw.sig->sig_mr);
1680         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1681
1682         if (ret) {
1683                 /*
1684                  * transport_generic_request_failure() expects to have
1685                  * plus two references to handle queue-full, so re-add
1686                  * one here as target-core will have already dropped
1687                  * it after the first isert_put_datain() callback.
1688                  */
1689                 kref_get(&cmd->cmd_kref);
1690                 transport_generic_request_failure(cmd, cmd->pi_err);
1691         } else {
1692                 /*
1693                  * XXX: isert_put_response() failure is not retried.
1694                  */
1695                 ret = isert_put_response(isert_conn->conn, isert_cmd->iscsi_cmd);
1696                 if (ret)
1697                         pr_warn_ratelimited("isert_put_response() ret: %d\n", ret);
1698         }
1699 }
1700
1701 static void
1702 isert_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1703 {
1704         struct isert_conn *isert_conn = wc->qp->qp_context;
1705         struct isert_device *device = isert_conn->device;
1706         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1707         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1708         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1709         struct se_cmd *se_cmd = &cmd->se_cmd;
1710         int ret = 0;
1711
1712         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1713                 isert_print_wc(wc, "rdma read");
1714                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1715                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1716                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1717                 return;
1718         }
1719
1720         isert_dbg("Cmd %p\n", isert_cmd);
1721
1722         iscsit_stop_dataout_timer(cmd);
1723
1724         if (isert_prot_cmd(isert_conn, se_cmd))
1725                 ret = isert_check_pi_status(se_cmd, isert_cmd->rw.sig->sig_mr);
1726         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1727         cmd->write_data_done = 0;
1728
1729         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1730         spin_lock_bh(&cmd->istate_lock);
1731         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1732         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1733         spin_unlock_bh(&cmd->istate_lock);
1734
1735         /*
1736          * transport_generic_request_failure() will drop the extra
1737          * se_cmd->cmd_kref reference after T10-PI error, and handle
1738          * any non-zero ->queue_status() callback error retries.
1739          */
1740         if (ret)
1741                 transport_generic_request_failure(se_cmd, se_cmd->pi_err);
1742         else
1743                 target_execute_cmd(se_cmd);
1744 }
1745
1746 static void
1747 isert_do_control_comp(struct work_struct *work)
1748 {
1749         struct isert_cmd *isert_cmd = container_of(work,
1750                         struct isert_cmd, comp_work);
1751         struct isert_conn *isert_conn = isert_cmd->conn;
1752         struct ib_device *ib_dev = isert_conn->cm_id->device;
1753         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1754
1755         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1756
1757         switch (cmd->i_state) {
1758         case ISTATE_SEND_TASKMGTRSP:
1759                 iscsit_tmr_post_handler(cmd, cmd->conn);
1760                 /* fall through */
1761         case ISTATE_SEND_REJECT:
1762         case ISTATE_SEND_TEXTRSP:
1763                 cmd->i_state = ISTATE_SENT_STATUS;
1764                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1765                                      ib_dev, false);
1766                 break;
1767         case ISTATE_SEND_LOGOUTRSP:
1768                 iscsit_logout_post_handler(cmd, cmd->conn);
1769                 break;
1770         default:
1771                 isert_err("Unknown i_state %d\n", cmd->i_state);
1772                 dump_stack();
1773                 break;
1774         }
1775 }
1776
1777 static void
1778 isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc)
1779 {
1780         struct isert_conn *isert_conn = wc->qp->qp_context;
1781         struct ib_device *ib_dev = isert_conn->cm_id->device;
1782         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1783
1784         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1785                 isert_print_wc(wc, "login send");
1786                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1787                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1788         }
1789
1790         isert_unmap_tx_desc(tx_desc, ib_dev);
1791 }
1792
1793 static void
1794 isert_send_done(struct ib_cq *cq, struct ib_wc *wc)
1795 {
1796         struct isert_conn *isert_conn = wc->qp->qp_context;
1797         struct ib_device *ib_dev = isert_conn->cm_id->device;
1798         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1799         struct isert_cmd *isert_cmd = tx_desc_to_cmd(tx_desc);
1800
1801         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1802                 isert_print_wc(wc, "send");
1803                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1804                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1805                 isert_completion_put(tx_desc, isert_cmd, ib_dev, true);
1806                 return;
1807         }
1808
1809         isert_dbg("Cmd %p\n", isert_cmd);
1810
1811         switch (isert_cmd->iscsi_cmd->i_state) {
1812         case ISTATE_SEND_TASKMGTRSP:
1813         case ISTATE_SEND_LOGOUTRSP:
1814         case ISTATE_SEND_REJECT:
1815         case ISTATE_SEND_TEXTRSP:
1816                 isert_unmap_tx_desc(tx_desc, ib_dev);
1817
1818                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1819                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1820                 return;
1821         default:
1822                 isert_cmd->iscsi_cmd->i_state = ISTATE_SENT_STATUS;
1823                 isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1824                 break;
1825         }
1826 }
1827
1828 static int
1829 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
1830 {
1831         struct ib_send_wr *wr_failed;
1832         int ret;
1833
1834         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
1835         if (ret) {
1836                 isert_err("ib_post_recv failed with %d\n", ret);
1837                 return ret;
1838         }
1839
1840         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
1841                            &wr_failed);
1842         if (ret) {
1843                 isert_err("ib_post_send failed with %d\n", ret);
1844                 return ret;
1845         }
1846         return ret;
1847 }
1848
1849 static int
1850 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1851 {
1852         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1853         struct isert_conn *isert_conn = conn->context;
1854         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1855         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
1856                                 &isert_cmd->tx_desc.iscsi_header;
1857
1858         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1859         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
1860         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1861         /*
1862          * Attach SENSE DATA payload to iSCSI Response PDU
1863          */
1864         if (cmd->se_cmd.sense_buffer &&
1865             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
1866             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
1867                 struct isert_device *device = isert_conn->device;
1868                 struct ib_device *ib_dev = device->ib_device;
1869                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
1870                 u32 padding, pdu_len;
1871
1872                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
1873                                    cmd->sense_buffer);
1874                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
1875
1876                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
1877                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
1878                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
1879
1880                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
1881                                 (void *)cmd->sense_buffer, pdu_len,
1882                                 DMA_TO_DEVICE);
1883                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
1884                         return -ENOMEM;
1885
1886                 isert_cmd->pdu_buf_len = pdu_len;
1887                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
1888                 tx_dsg->length  = pdu_len;
1889                 tx_dsg->lkey    = device->pd->local_dma_lkey;
1890                 isert_cmd->tx_desc.num_sge = 2;
1891         }
1892
1893         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1894
1895         isert_dbg("Posting SCSI Response\n");
1896
1897         return isert_post_response(isert_conn, isert_cmd);
1898 }
1899
1900 static void
1901 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1902 {
1903         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1904         struct isert_conn *isert_conn = conn->context;
1905
1906         spin_lock_bh(&conn->cmd_lock);
1907         if (!list_empty(&cmd->i_conn_node))
1908                 list_del_init(&cmd->i_conn_node);
1909         spin_unlock_bh(&conn->cmd_lock);
1910
1911         if (cmd->data_direction == DMA_TO_DEVICE)
1912                 iscsit_stop_dataout_timer(cmd);
1913         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1914 }
1915
1916 static enum target_prot_op
1917 isert_get_sup_prot_ops(struct iscsi_conn *conn)
1918 {
1919         struct isert_conn *isert_conn = conn->context;
1920         struct isert_device *device = isert_conn->device;
1921
1922         if (conn->tpg->tpg_attrib.t10_pi) {
1923                 if (device->pi_capable) {
1924                         isert_info("conn %p PI offload enabled\n", isert_conn);
1925                         isert_conn->pi_support = true;
1926                         return TARGET_PROT_ALL;
1927                 }
1928         }
1929
1930         isert_info("conn %p PI offload disabled\n", isert_conn);
1931         isert_conn->pi_support = false;
1932
1933         return TARGET_PROT_NORMAL;
1934 }
1935
1936 static int
1937 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
1938                 bool nopout_response)
1939 {
1940         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1941         struct isert_conn *isert_conn = conn->context;
1942         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1943
1944         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1945         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
1946                                &isert_cmd->tx_desc.iscsi_header,
1947                                nopout_response);
1948         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1949         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1950
1951         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
1952
1953         return isert_post_response(isert_conn, isert_cmd);
1954 }
1955
1956 static int
1957 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1958 {
1959         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1960         struct isert_conn *isert_conn = conn->context;
1961         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1962
1963         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1964         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
1965                                 &isert_cmd->tx_desc.iscsi_header);
1966         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1967         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1968
1969         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
1970
1971         return isert_post_response(isert_conn, isert_cmd);
1972 }
1973
1974 static int
1975 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1976 {
1977         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1978         struct isert_conn *isert_conn = conn->context;
1979         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1980
1981         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1982         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
1983                                   &isert_cmd->tx_desc.iscsi_header);
1984         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1985         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1986
1987         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
1988
1989         return isert_post_response(isert_conn, isert_cmd);
1990 }
1991
1992 static int
1993 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1994 {
1995         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1996         struct isert_conn *isert_conn = conn->context;
1997         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1998         struct isert_device *device = isert_conn->device;
1999         struct ib_device *ib_dev = device->ib_device;
2000         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2001         struct iscsi_reject *hdr =
2002                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2003
2004         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2005         iscsit_build_reject(cmd, conn, hdr);
2006         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2007
2008         hton24(hdr->dlength, ISCSI_HDR_LEN);
2009         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2010                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2011                         DMA_TO_DEVICE);
2012         if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2013                 return -ENOMEM;
2014         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2015         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2016         tx_dsg->length  = ISCSI_HDR_LEN;
2017         tx_dsg->lkey    = device->pd->local_dma_lkey;
2018         isert_cmd->tx_desc.num_sge = 2;
2019
2020         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2021
2022         isert_dbg("conn %p Posting Reject\n", isert_conn);
2023
2024         return isert_post_response(isert_conn, isert_cmd);
2025 }
2026
2027 static int
2028 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2029 {
2030         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2031         struct isert_conn *isert_conn = conn->context;
2032         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2033         struct iscsi_text_rsp *hdr =
2034                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2035         u32 txt_rsp_len;
2036         int rc;
2037
2038         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2039         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2040         if (rc < 0)
2041                 return rc;
2042
2043         txt_rsp_len = rc;
2044         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2045
2046         if (txt_rsp_len) {
2047                 struct isert_device *device = isert_conn->device;
2048                 struct ib_device *ib_dev = device->ib_device;
2049                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2050                 void *txt_rsp_buf = cmd->buf_ptr;
2051
2052                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2053                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2054                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2055                         return -ENOMEM;
2056
2057                 isert_cmd->pdu_buf_len = txt_rsp_len;
2058                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2059                 tx_dsg->length  = txt_rsp_len;
2060                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2061                 isert_cmd->tx_desc.num_sge = 2;
2062         }
2063         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2064
2065         isert_dbg("conn %p Text Response\n", isert_conn);
2066
2067         return isert_post_response(isert_conn, isert_cmd);
2068 }
2069
2070 static inline void
2071 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2072                      struct ib_sig_domain *domain)
2073 {
2074         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2075         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2076         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2077         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2078         /*
2079          * At the moment we hard code those, but if in the future
2080          * the target core would like to use it, we will take it
2081          * from se_cmd.
2082          */
2083         domain->sig.dif.apptag_check_mask = 0xffff;
2084         domain->sig.dif.app_escape = true;
2085         domain->sig.dif.ref_escape = true;
2086         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2087             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2088                 domain->sig.dif.ref_remap = true;
2089 };
2090
2091 static int
2092 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2093 {
2094         memset(sig_attrs, 0, sizeof(*sig_attrs));
2095
2096         switch (se_cmd->prot_op) {
2097         case TARGET_PROT_DIN_INSERT:
2098         case TARGET_PROT_DOUT_STRIP:
2099                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2100                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2101                 break;
2102         case TARGET_PROT_DOUT_INSERT:
2103         case TARGET_PROT_DIN_STRIP:
2104                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2105                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2106                 break;
2107         case TARGET_PROT_DIN_PASS:
2108         case TARGET_PROT_DOUT_PASS:
2109                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2110                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2111                 break;
2112         default:
2113                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2114                 return -EINVAL;
2115         }
2116
2117         if (se_cmd->prot_checks & TARGET_DIF_CHECK_GUARD)
2118                 sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
2119         if (se_cmd->prot_checks & TARGET_DIF_CHECK_APPTAG)
2120                 sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
2121         if (se_cmd->prot_checks & TARGET_DIF_CHECK_REFTAG)
2122                 sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
2123
2124         return 0;
2125 }
2126
2127 static int
2128 isert_rdma_rw_ctx_post(struct isert_cmd *cmd, struct isert_conn *conn,
2129                 struct ib_cqe *cqe, struct ib_send_wr *chain_wr)
2130 {
2131         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
2132         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
2133         u8 port_num = conn->cm_id->port_num;
2134         u64 addr;
2135         u32 rkey, offset;
2136         int ret;
2137
2138         if (cmd->ctx_init_done)
2139                 goto rdma_ctx_post;
2140
2141         if (dir == DMA_FROM_DEVICE) {
2142                 addr = cmd->write_va;
2143                 rkey = cmd->write_stag;
2144                 offset = cmd->iscsi_cmd->write_data_done;
2145         } else {
2146                 addr = cmd->read_va;
2147                 rkey = cmd->read_stag;
2148                 offset = 0;
2149         }
2150
2151         if (isert_prot_cmd(conn, se_cmd)) {
2152                 struct ib_sig_attrs sig_attrs;
2153
2154                 ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2155                 if (ret)
2156                         return ret;
2157
2158                 WARN_ON_ONCE(offset);
2159                 ret = rdma_rw_ctx_signature_init(&cmd->rw, conn->qp, port_num,
2160                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2161                                 se_cmd->t_prot_sg, se_cmd->t_prot_nents,
2162                                 &sig_attrs, addr, rkey, dir);
2163         } else {
2164                 ret = rdma_rw_ctx_init(&cmd->rw, conn->qp, port_num,
2165                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2166                                 offset, addr, rkey, dir);
2167         }
2168
2169         if (ret < 0) {
2170                 isert_err("Cmd: %p failed to prepare RDMA res\n", cmd);
2171                 return ret;
2172         }
2173
2174         cmd->ctx_init_done = true;
2175
2176 rdma_ctx_post:
2177         ret = rdma_rw_ctx_post(&cmd->rw, conn->qp, port_num, cqe, chain_wr);
2178         if (ret < 0)
2179                 isert_err("Cmd: %p failed to post RDMA res\n", cmd);
2180         return ret;
2181 }
2182
2183 static int
2184 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2185 {
2186         struct se_cmd *se_cmd = &cmd->se_cmd;
2187         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2188         struct isert_conn *isert_conn = conn->context;
2189         struct ib_cqe *cqe = NULL;
2190         struct ib_send_wr *chain_wr = NULL;
2191         int rc;
2192
2193         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2194                  isert_cmd, se_cmd->data_length);
2195
2196         if (isert_prot_cmd(isert_conn, se_cmd)) {
2197                 isert_cmd->tx_desc.tx_cqe.done = isert_rdma_write_done;
2198                 cqe = &isert_cmd->tx_desc.tx_cqe;
2199         } else {
2200                 /*
2201                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2202                  */
2203                 isert_create_send_desc(isert_conn, isert_cmd,
2204                                        &isert_cmd->tx_desc);
2205                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2206                                      &isert_cmd->tx_desc.iscsi_header);
2207                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2208                 isert_init_send_wr(isert_conn, isert_cmd,
2209                                    &isert_cmd->tx_desc.send_wr);
2210
2211                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2212                 if (rc) {
2213                         isert_err("ib_post_recv failed with %d\n", rc);
2214                         return rc;
2215                 }
2216
2217                 chain_wr = &isert_cmd->tx_desc.send_wr;
2218         }
2219
2220         rc = isert_rdma_rw_ctx_post(isert_cmd, isert_conn, cqe, chain_wr);
2221         isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ rc: %d\n",
2222                   isert_cmd, rc);
2223         return rc;
2224 }
2225
2226 static int
2227 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2228 {
2229         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2230         int ret;
2231
2232         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2233                  isert_cmd, cmd->se_cmd.data_length, cmd->write_data_done);
2234
2235         isert_cmd->tx_desc.tx_cqe.done = isert_rdma_read_done;
2236         ret = isert_rdma_rw_ctx_post(isert_cmd, conn->context,
2237                                      &isert_cmd->tx_desc.tx_cqe, NULL);
2238
2239         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE rc: %d\n",
2240                  isert_cmd, ret);
2241         return ret;
2242 }
2243
2244 static int
2245 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2246 {
2247         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2248         int ret = 0;
2249
2250         switch (state) {
2251         case ISTATE_REMOVE:
2252                 spin_lock_bh(&conn->cmd_lock);
2253                 list_del_init(&cmd->i_conn_node);
2254                 spin_unlock_bh(&conn->cmd_lock);
2255                 isert_put_cmd(isert_cmd, true);
2256                 break;
2257         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2258                 ret = isert_put_nopin(cmd, conn, false);
2259                 break;
2260         default:
2261                 isert_err("Unknown immediate state: 0x%02x\n", state);
2262                 ret = -EINVAL;
2263                 break;
2264         }
2265
2266         return ret;
2267 }
2268
2269 static int
2270 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2271 {
2272         struct isert_conn *isert_conn = conn->context;
2273         int ret;
2274
2275         switch (state) {
2276         case ISTATE_SEND_LOGOUTRSP:
2277                 ret = isert_put_logout_rsp(cmd, conn);
2278                 if (!ret)
2279                         isert_conn->logout_posted = true;
2280                 break;
2281         case ISTATE_SEND_NOPIN:
2282                 ret = isert_put_nopin(cmd, conn, true);
2283                 break;
2284         case ISTATE_SEND_TASKMGTRSP:
2285                 ret = isert_put_tm_rsp(cmd, conn);
2286                 break;
2287         case ISTATE_SEND_REJECT:
2288                 ret = isert_put_reject(cmd, conn);
2289                 break;
2290         case ISTATE_SEND_TEXTRSP:
2291                 ret = isert_put_text_rsp(cmd, conn);
2292                 break;
2293         case ISTATE_SEND_STATUS:
2294                 /*
2295                  * Special case for sending non GOOD SCSI status from TX thread
2296                  * context during pre se_cmd excecution failure.
2297                  */
2298                 ret = isert_put_response(conn, cmd);
2299                 break;
2300         default:
2301                 isert_err("Unknown response state: 0x%02x\n", state);
2302                 ret = -EINVAL;
2303                 break;
2304         }
2305
2306         return ret;
2307 }
2308
2309 struct rdma_cm_id *
2310 isert_setup_id(struct isert_np *isert_np)
2311 {
2312         struct iscsi_np *np = isert_np->np;
2313         struct rdma_cm_id *id;
2314         struct sockaddr *sa;
2315         int ret;
2316
2317         sa = (struct sockaddr *)&np->np_sockaddr;
2318         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
2319
2320         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
2321                             RDMA_PS_TCP, IB_QPT_RC);
2322         if (IS_ERR(id)) {
2323                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
2324                 ret = PTR_ERR(id);
2325                 goto out;
2326         }
2327         isert_dbg("id %p context %p\n", id, id->context);
2328
2329         ret = rdma_bind_addr(id, sa);
2330         if (ret) {
2331                 isert_err("rdma_bind_addr() failed: %d\n", ret);
2332                 goto out_id;
2333         }
2334
2335         ret = rdma_listen(id, 0);
2336         if (ret) {
2337                 isert_err("rdma_listen() failed: %d\n", ret);
2338                 goto out_id;
2339         }
2340
2341         return id;
2342 out_id:
2343         rdma_destroy_id(id);
2344 out:
2345         return ERR_PTR(ret);
2346 }
2347
2348 static int
2349 isert_setup_np(struct iscsi_np *np,
2350                struct sockaddr_storage *ksockaddr)
2351 {
2352         struct isert_np *isert_np;
2353         struct rdma_cm_id *isert_lid;
2354         int ret;
2355
2356         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
2357         if (!isert_np)
2358                 return -ENOMEM;
2359
2360         sema_init(&isert_np->sem, 0);
2361         mutex_init(&isert_np->mutex);
2362         INIT_LIST_HEAD(&isert_np->accepted);
2363         INIT_LIST_HEAD(&isert_np->pending);
2364         isert_np->np = np;
2365
2366         /*
2367          * Setup the np->np_sockaddr from the passed sockaddr setup
2368          * in iscsi_target_configfs.c code..
2369          */
2370         memcpy(&np->np_sockaddr, ksockaddr,
2371                sizeof(struct sockaddr_storage));
2372
2373         isert_lid = isert_setup_id(isert_np);
2374         if (IS_ERR(isert_lid)) {
2375                 ret = PTR_ERR(isert_lid);
2376                 goto out;
2377         }
2378
2379         isert_np->cm_id = isert_lid;
2380         np->np_context = isert_np;
2381
2382         return 0;
2383
2384 out:
2385         kfree(isert_np);
2386
2387         return ret;
2388 }
2389
2390 static int
2391 isert_rdma_accept(struct isert_conn *isert_conn)
2392 {
2393         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2394         struct rdma_conn_param cp;
2395         int ret;
2396         struct iser_cm_hdr rsp_hdr;
2397
2398         memset(&cp, 0, sizeof(struct rdma_conn_param));
2399         cp.initiator_depth = isert_conn->initiator_depth;
2400         cp.retry_count = 7;
2401         cp.rnr_retry_count = 7;
2402
2403         memset(&rsp_hdr, 0, sizeof(rsp_hdr));
2404         rsp_hdr.flags = ISERT_ZBVA_NOT_USED;
2405         if (!isert_conn->snd_w_inv)
2406                 rsp_hdr.flags = rsp_hdr.flags | ISERT_SEND_W_INV_NOT_USED;
2407         cp.private_data = (void *)&rsp_hdr;
2408         cp.private_data_len = sizeof(rsp_hdr);
2409
2410         ret = rdma_accept(cm_id, &cp);
2411         if (ret) {
2412                 isert_err("rdma_accept() failed with: %d\n", ret);
2413                 return ret;
2414         }
2415
2416         return 0;
2417 }
2418
2419 static int
2420 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
2421 {
2422         struct isert_conn *isert_conn = conn->context;
2423         int ret;
2424
2425         isert_info("before login_req comp conn: %p\n", isert_conn);
2426         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
2427         if (ret) {
2428                 isert_err("isert_conn %p interrupted before got login req\n",
2429                           isert_conn);
2430                 return ret;
2431         }
2432         reinit_completion(&isert_conn->login_req_comp);
2433
2434         /*
2435          * For login requests after the first PDU, isert_rx_login_req() will
2436          * kick schedule_delayed_work(&conn->login_work) as the packet is
2437          * received, which turns this callback from iscsi_target_do_login_rx()
2438          * into a NOP.
2439          */
2440         if (!login->first_request)
2441                 return 0;
2442
2443         isert_rx_login_req(isert_conn);
2444
2445         isert_info("before login_comp conn: %p\n", conn);
2446         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
2447         if (ret)
2448                 return ret;
2449
2450         isert_info("processing login->req: %p\n", login->req);
2451
2452         return 0;
2453 }
2454
2455 static void
2456 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
2457                     struct isert_conn *isert_conn)
2458 {
2459         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2460         struct rdma_route *cm_route = &cm_id->route;
2461
2462         conn->login_family = np->np_sockaddr.ss_family;
2463
2464         conn->login_sockaddr = cm_route->addr.dst_addr;
2465         conn->local_sockaddr = cm_route->addr.src_addr;
2466 }
2467
2468 static int
2469 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
2470 {
2471         struct isert_np *isert_np = np->np_context;
2472         struct isert_conn *isert_conn;
2473         int ret;
2474
2475 accept_wait:
2476         ret = down_interruptible(&isert_np->sem);
2477         if (ret)
2478                 return -ENODEV;
2479
2480         spin_lock_bh(&np->np_thread_lock);
2481         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
2482                 spin_unlock_bh(&np->np_thread_lock);
2483                 isert_dbg("np_thread_state %d\n",
2484                          np->np_thread_state);
2485                 /**
2486                  * No point in stalling here when np_thread
2487                  * is in state RESET/SHUTDOWN/EXIT - bail
2488                  **/
2489                 return -ENODEV;
2490         }
2491         spin_unlock_bh(&np->np_thread_lock);
2492
2493         mutex_lock(&isert_np->mutex);
2494         if (list_empty(&isert_np->pending)) {
2495                 mutex_unlock(&isert_np->mutex);
2496                 goto accept_wait;
2497         }
2498         isert_conn = list_first_entry(&isert_np->pending,
2499                         struct isert_conn, node);
2500         list_del_init(&isert_conn->node);
2501         mutex_unlock(&isert_np->mutex);
2502
2503         conn->context = isert_conn;
2504         isert_conn->conn = conn;
2505         isert_conn->state = ISER_CONN_BOUND;
2506
2507         isert_set_conn_info(np, conn, isert_conn);
2508
2509         isert_dbg("Processing isert_conn: %p\n", isert_conn);
2510
2511         return 0;
2512 }
2513
2514 static void
2515 isert_free_np(struct iscsi_np *np)
2516 {
2517         struct isert_np *isert_np = np->np_context;
2518         struct isert_conn *isert_conn, *n;
2519
2520         if (isert_np->cm_id)
2521                 rdma_destroy_id(isert_np->cm_id);
2522
2523         /*
2524          * FIXME: At this point we don't have a good way to insure
2525          * that at this point we don't have hanging connections that
2526          * completed RDMA establishment but didn't start iscsi login
2527          * process. So work-around this by cleaning up what ever piled
2528          * up in accepted and pending lists.
2529          */
2530         mutex_lock(&isert_np->mutex);
2531         if (!list_empty(&isert_np->pending)) {
2532                 isert_info("Still have isert pending connections\n");
2533                 list_for_each_entry_safe(isert_conn, n,
2534                                          &isert_np->pending,
2535                                          node) {
2536                         isert_info("cleaning isert_conn %p state (%d)\n",
2537                                    isert_conn, isert_conn->state);
2538                         isert_connect_release(isert_conn);
2539                 }
2540         }
2541
2542         if (!list_empty(&isert_np->accepted)) {
2543                 isert_info("Still have isert accepted connections\n");
2544                 list_for_each_entry_safe(isert_conn, n,
2545                                          &isert_np->accepted,
2546                                          node) {
2547                         isert_info("cleaning isert_conn %p state (%d)\n",
2548                                    isert_conn, isert_conn->state);
2549                         isert_connect_release(isert_conn);
2550                 }
2551         }
2552         mutex_unlock(&isert_np->mutex);
2553
2554         np->np_context = NULL;
2555         kfree(isert_np);
2556 }
2557
2558 static void isert_release_work(struct work_struct *work)
2559 {
2560         struct isert_conn *isert_conn = container_of(work,
2561                                                      struct isert_conn,
2562                                                      release_work);
2563
2564         isert_info("Starting release conn %p\n", isert_conn);
2565
2566         mutex_lock(&isert_conn->mutex);
2567         isert_conn->state = ISER_CONN_DOWN;
2568         mutex_unlock(&isert_conn->mutex);
2569
2570         isert_info("Destroying conn %p\n", isert_conn);
2571         isert_put_conn(isert_conn);
2572 }
2573
2574 static void
2575 isert_wait4logout(struct isert_conn *isert_conn)
2576 {
2577         struct iscsi_conn *conn = isert_conn->conn;
2578
2579         isert_info("conn %p\n", isert_conn);
2580
2581         if (isert_conn->logout_posted) {
2582                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
2583                 wait_for_completion_timeout(&conn->conn_logout_comp,
2584                                             SECONDS_FOR_LOGOUT_COMP * HZ);
2585         }
2586 }
2587
2588 static void
2589 isert_wait4cmds(struct iscsi_conn *conn)
2590 {
2591         isert_info("iscsi_conn %p\n", conn);
2592
2593         if (conn->sess) {
2594                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
2595                 target_wait_for_sess_cmds(conn->sess->se_sess);
2596         }
2597 }
2598
2599 /**
2600  * isert_put_unsol_pending_cmds() - Drop commands waiting for
2601  *     unsolicitate dataout
2602  * @conn:    iscsi connection
2603  *
2604  * We might still have commands that are waiting for unsolicited
2605  * dataouts messages. We must put the extra reference on those
2606  * before blocking on the target_wait_for_session_cmds
2607  */
2608 static void
2609 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
2610 {
2611         struct iscsi_cmd *cmd, *tmp;
2612         static LIST_HEAD(drop_cmd_list);
2613
2614         spin_lock_bh(&conn->cmd_lock);
2615         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
2616                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
2617                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
2618                     (cmd->write_data_done < cmd->se_cmd.data_length))
2619                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
2620         }
2621         spin_unlock_bh(&conn->cmd_lock);
2622
2623         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
2624                 list_del_init(&cmd->i_conn_node);
2625                 if (cmd->i_state != ISTATE_REMOVE) {
2626                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2627
2628                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
2629                         isert_put_cmd(isert_cmd, true);
2630                 }
2631         }
2632 }
2633
2634 static void isert_wait_conn(struct iscsi_conn *conn)
2635 {
2636         struct isert_conn *isert_conn = conn->context;
2637
2638         isert_info("Starting conn %p\n", isert_conn);
2639
2640         mutex_lock(&isert_conn->mutex);
2641         isert_conn_terminate(isert_conn);
2642         mutex_unlock(&isert_conn->mutex);
2643
2644         ib_drain_qp(isert_conn->qp);
2645         isert_put_unsol_pending_cmds(conn);
2646         isert_wait4cmds(conn);
2647         isert_wait4logout(isert_conn);
2648
2649         queue_work(isert_release_wq, &isert_conn->release_work);
2650 }
2651
2652 static void isert_free_conn(struct iscsi_conn *conn)
2653 {
2654         struct isert_conn *isert_conn = conn->context;
2655
2656         ib_drain_qp(isert_conn->qp);
2657         isert_put_conn(isert_conn);
2658 }
2659
2660 static void isert_get_rx_pdu(struct iscsi_conn *conn)
2661 {
2662         struct completion comp;
2663
2664         init_completion(&comp);
2665
2666         wait_for_completion_interruptible(&comp);
2667 }
2668
2669 static struct iscsit_transport iser_target_transport = {
2670         .name                   = "IB/iSER",
2671         .transport_type         = ISCSI_INFINIBAND,
2672         .rdma_shutdown          = true,
2673         .priv_size              = sizeof(struct isert_cmd),
2674         .owner                  = THIS_MODULE,
2675         .iscsit_setup_np        = isert_setup_np,
2676         .iscsit_accept_np       = isert_accept_np,
2677         .iscsit_free_np         = isert_free_np,
2678         .iscsit_wait_conn       = isert_wait_conn,
2679         .iscsit_free_conn       = isert_free_conn,
2680         .iscsit_get_login_rx    = isert_get_login_rx,
2681         .iscsit_put_login_tx    = isert_put_login_tx,
2682         .iscsit_immediate_queue = isert_immediate_queue,
2683         .iscsit_response_queue  = isert_response_queue,
2684         .iscsit_get_dataout     = isert_get_dataout,
2685         .iscsit_queue_data_in   = isert_put_datain,
2686         .iscsit_queue_status    = isert_put_response,
2687         .iscsit_aborted_task    = isert_aborted_task,
2688         .iscsit_get_rx_pdu      = isert_get_rx_pdu,
2689         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
2690 };
2691
2692 static int __init isert_init(void)
2693 {
2694         int ret;
2695
2696         isert_comp_wq = alloc_workqueue("isert_comp_wq",
2697                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
2698         if (!isert_comp_wq) {
2699                 isert_err("Unable to allocate isert_comp_wq\n");
2700                 return -ENOMEM;
2701         }
2702
2703         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
2704                                         WQ_UNBOUND_MAX_ACTIVE);
2705         if (!isert_release_wq) {
2706                 isert_err("Unable to allocate isert_release_wq\n");
2707                 ret = -ENOMEM;
2708                 goto destroy_comp_wq;
2709         }
2710
2711         iscsit_register_transport(&iser_target_transport);
2712         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
2713
2714         return 0;
2715
2716 destroy_comp_wq:
2717         destroy_workqueue(isert_comp_wq);
2718
2719         return ret;
2720 }
2721
2722 static void __exit isert_exit(void)
2723 {
2724         flush_scheduled_work();
2725         destroy_workqueue(isert_release_wq);
2726         destroy_workqueue(isert_comp_wq);
2727         iscsit_unregister_transport(&iser_target_transport);
2728         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
2729 }
2730
2731 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
2732 MODULE_AUTHOR("nab@Linux-iSCSI.org");
2733 MODULE_LICENSE("GPL");
2734
2735 module_init(isert_init);
2736 module_exit(isert_exit);